A NNALI DELLA
S CUOLA N ORMALE S UPERIORE DI P ISA Classe di Scienze
N ICOLAI V. K RYLOV
Fully nonlinear second order elliptic equations : recent development
Annali della Scuola Normale Superiore di Pisa, Classe di Scienze 4
esérie, tome 25, n
o3-4 (1997), p. 569-595
<http://www.numdam.org/item?id=ASNSP_1997_4_25_3-4_569_0>
© Scuola Normale Superiore, Pisa, 1997, tous droits réservés.
L’accès aux archives de la revue « Annali della Scuola Normale Superiore di Pisa, Classe di Scienze » (http://www.sns.it/it/edizioni/riviste/annaliscienze/) implique l’accord avec les conditions générales d’utilisation (http://www.numdam.org/conditions). Toute utilisa- tion commerciale ou impression systématique est constitutive d’une infraction pénale.
Toute copie ou impression de ce fichier doit contenir la présente mention de copyright.
Article numérisé dans le cadre du programme Numérisation de documents anciens mathématiques
http://www.numdam.org/
569
Fully Nonlinear Second Order Elliptic Equations:
Recent Development
NICOLAI V. KRYLOV
Abstract. A short discussion of the history of the theory of fully nonlinear second-
order elliptic equations is presented starting with the beginning of the century.
Then an account of the explosion of results during the last decade is given. This explosion is based entirely on a generalization for nondivergence form linear operators of the celebrated De Giorgi result bearing an Holder continuity. This is
an extended version of a 1.5 hour talk at Mathfest, Burlington, Vermont, Aug 6- 8, 1995.
Mathematics Subject Classification (1991): 35J60.
1. - Introduction
It seems Bernstein
[8]10
in 1910 was the first to introducegeneral
methodsof
solving
nonlinearelliptic equations.
These are: the method ofcontinuity
andthe method of a
priori
estimates. He consideredequations
with twoindependent
variables and showed that for
proving
thesolvability
of suchequations
it sufficesto establish a
priori
estimates for absolute values of the first two derivatives of solutions. For Bernstein theequation
where Q
is
a domain inR~,
should be calledelliptic
ifthe
matrix(F~l~ )
isdefinite. This definition is
quite
natural when theequation
is linear with respectto second order
derivatives,
but is not of muchhelp
forfully
nonlinearequations.
For
example,
before 1910 in 1903 Minkowski[65]°3 proved
existence anduniqueness
of convex surface withprescribed
Gaussian curvature in Euclideanspace. He did not prove however that this surface is from class
C2.
Analytically,
the Minkowskiproblem
involvessolving
ahighly
nonlinearpartial
differentialequation
ofMonge-Ampere
type. Anexample
of such anThis work was supported in part by NSF grant DMS-9302516.
(1997), pp. 569-595
equation,
for a convex functionu (x) - u (x 1, ... , xd )
defined in a domain inR d
is thefollowing simplest Monge-Ampere equation
where
f (x )
is agiven
function. Thisequation
is notelliptic
in the sense ofBernstein.
It turns out that
(1.2)
and similarequations
can be well understood even fornon differentiable
functions,
so that one caninvestigate generalized
solutions.This was done
by
Aleksandrov[ 1 ]5g
in 1958 and led to some remarkable results for linearequation
which 20 years later turned out to constitute the basis of thegeneral theory
offully
nonlinearelliptic equations.
We mean the so-called Aleksandrov maximumprinciple
and Aleksandrov estimates(see [2]61
and[3]63).
The smoothness of Minkowski’s two-dimensional
generalized
solution wasstudied
by
well-knownmathematicians,
such asLewy [59]3~,
Miranda[66]39,
9Pogorelov [69]52, Nirenberg [67]53,
Calabi[19]58 ,
Bakel’man[6]65.
In 1971that is 68 years after Minkowski’s
work, Pogorelov
in[71 ] ~
1finally proved
thatthe solution is indeed from class
C2
in multidimensional case. Nevertheless hisproofs
in[71]71
1 and[72]~5
contained what looked like a vicious circle andonly
in 1977
Cheng
and Yau[2 1]77
showed how to avoid it.To prove the existence of solutions of
equations
like(1.2) by
the methodsknown before 1981 was no easy task. It involved
finding
apriori
estimatesfor solutions and their derivatives up to the third order.
Big part
of the work is based ondifferentiating (1.2)
three times and on certainextremely cleverly organized manipulations
inventedby
Calabi. After 1981 theapproach
tofully
nonlinear
equations changed dramatically.
We discuss this in Section 3.Until 1971 the
theory
offully
nonlinearelliptic equations only
consisted ofthe
theory of Monge-Ampere equations.
In 1971appeared
the so-called Bellmanequations
which stemmed from thetheory
of controlled diffusion processes. Thetypical representative
of Bellman’sequations
is thefollowing:
where A is a set and
(aij)
=(a’j)* >
0.To non
specialists equations
in form(1.3) might
look artificial.Indeed, equations
which arise in other fields of mathematics look different. Forexample,
in many natural
geometrical applications
there appearequations
like theMonge- Amp6re equations
andprescribed
curvatureequations
or more
general equations
where
Hyn
denotes the m-thsymmetric polynomial
of the curvature matrix ofthe
graph
of u.Equations
of the form of thecomplex Monge-Ampere equation:
are also very
popular
in the literature. Theseequations
arise incomplex
geom- etry andcomplex analysis.
Thefollowing equation
where I is the d x d unit
matrix,
comes from thetheory
of calibratedgeometries (see Caffarelli, Nirenberg
andSpruck [16]85)
in connection withabsolutely volume-minimizing
submanifolds ofR~.
Itsparticular
case(when
d =3)
was considered earlier
by Pogorelov [70]6°
in two dimensions andby
the au-thor
[42] g3 .
It is worthnoting
thatalthough equations ( 1.4)-( 1.7)
look different from the Bellmanequation (1.3),
nevertheless each of them isequivalent
to acorresponding
Bellmanequation
of type(1.3).
Withregard
to(1.4)
and(1.6)
this fact was known in
1971,
which made thetheory
of Bellman’sequations
very attractive.Really general theory
offully
nonlinearelliptic equations emerged
from the
theory
of Bellman’sequations.
We will see later that even now themost
general theory
offully
nonlinearelliptic equations,
infact,
reduces to thetheory
of Bellman’sequations.
At
early stages
of thetheory
of Bellman’sequations
theonly
availablemethods were those of the
theory
ofprobability.
It is remarkable that results obtainedby
these methods aresharp
in manysituations,
and even now some ofthem,
which admitpurely analytic formulation,
areonly
obtainedby probabilistic
means. The reader can make an
acquaintance
with thecorresponding
resultsstarting
with the bookby Fleming
and Sooner[27]93
and with references thereinto which we
only
addPragarauskas [73]82
andKrylov [46]g9.
By
the way for one of discontinuous controlled processes considered in[73] 82
Bellman’s
equation
takes thefollowing
form:Until 1982 the
probabilistic
methods were the mostpowerful
in thegeneral theory
offully
nonlinearelliptic equations.
The situationchanged dramatically
in 1982 when Evans
[25]g2
and the author[40]g2 proved
thesolvability
inc2+a of a
broad class of Bellman’selliptic
andparabolic equations.
Theproofs
were based on the
theory
of linearequations and,
inparticular,
on the factthat one can estimate the Holder constant for solutions of linear
equations
withmeasurable coefficients. The latter fact was
previously
established in 1979by
Safonov and the author
[52]~9 (see
also[53]g°
and a beautifulexposition by
Safonov for
elliptic
case in[75]80) ,
as mentioned before on the sole basis of Aleksandrov’s estimates.Remarkably enough, although
theproofs
in[53]g°
and
[75]80
are written in PDE terms, allunderlying
ideas areprobabilistic,
andperhaps
this was the reasonwhy people
from PDEs did not succeed inobtaining
the estimate before. After this the
general
PDEtheory
offully
nonlinearelliptic equations
started up, and below we willgive
areport
onmajor developments
of this
theory.
The article is
organized
as follows. In Section 2 wegive
ageneral
notionof
fully
nonlinearelliptic equation
and some existence theorems. Section 3 is devoted to resultsbearing
on thegeneral theory
offully
nonlinearuniformly elliptic equations
and Section 4 contains a discussion of results for thegeneral theory
offully
nonlineardegenerate elliptic equations.
In Section 5 wespeak
about
equations
related to theMonge-Ampere equations.
Section 6 containsa new
(and probably
thefirst)
result on the rate of convergence of numericalapproximations
forfully
nonlineardegenerate elliptic equations.
We havealready
mentioned above that the modem
theory
offully
nonlinearelliptic equations
is based
entirely
on somedeep
results from lineartheory.
Also many new brilliant ideas andtechniques appeared.
In Section 7 we present one of them which is Safonov’sproof
of the Holder-Korn-Lichtenstein-Giraud estimate for theLaplacian.
Thisproof
wasdesigned
for nonlinearequations
and turned outto be shorter and easier than usual ones even for the
simplest
linearequation.
Inmy
opinion
hisproof
should be part ofgeneral
mathematical education.Finally,
not as brilliant and not a very
popular
technical idea ispresented
in Section 8.Exploiting
this idea allowed the author toget
some verygeneral
results onfully
nonlinear
degenerate elliptic equations.
Also the idea is of ageneral
characterand
might
be of interest to mathematicians from other areas.It is to be said that the literature on
fully
nonlinearelliptic equations
isreally immense,
we present here a report on theonly
part of it which is closeto interests of the author. In
particular,
we do not discuss concreteapplications
in which
equations
we discuss arose.2. - A
general
notion offully
nonlinearelliptic equation
andexamples
Conceivably,
the firstquestion
which arises when atheory
starts is: whatis the main
object
ofinvestigation? Interestingly enough
this was not the firstquestion
addressed in the case of thetheory
offully
nonlinearelliptic equations.
The reason for this is that there were
enough
oldproblems regarding fully
nonlinear
equations
which came up earlier andthey
were to be solved in the firstplace.
Now when the
general theory
is rather welldeveloped,
one may think howto make the field of its
applications
as wide aspossible
and the number ofpeople
who can use it aslarge
aspossible.
We have the
following
situation. From the onehand,
ahuge variety
ofresults is available in the
theory.
On the other handhowever,
it turns out that ifan
inexperienced
reader meets afully
nonlinear second orderpartial
differentialequation
in hisinvestigations
and tries to get any informationconcerning
itssolvability
from theliterature,
then almostcertainly
he fails to find what heneeds,
unless he considers anequation
that isexactly
one which hadalready
been treated. The
point
is that in thegeneral theory
we treatonly equations
which
satisfy
certain conditions and whileconsidering examples,
we show howto
transform
theequations
in theseexamples
to otherequations
to which thetheory
isapplicable. Therefore,
from thepoint
of view ofapplications
the mainquestion
is how to describe insimple
terms the mostgeneral
situation when onecan make an
appropriate
transformation. In other terms, one needs ageneral
notion of
fully
nonlinearelliptic equation.
Naturally,
the type ofequation
should be definedonly by
the way of de-pendence
of F onD2 u,
that is we call ourequation (1.1) elliptic
if forany p E y E S2 and z E R the
following equation
in Q iselliptic:
F(D2u(x),
p, Z,y)
= 0.Therefore,
we have to concentrate on the case when Fdepends only
on the matrix of second order derivatives of u, in otherwords,
we have to consider the
equation
We assume of course that
Usually
in the literature on nonlinearelliptic equations (see,
forinstance,
[22]62, [29]83, [14]g4, [15]gs, [44]g5)
one accepts the definitionby
Bernsteinand
equation (2.8)
is calledelliptic
if the matrix isnonnegative (or nonpositive)
for all arguments. As we have noticedabove,
this excludes atonce even the
simplest Monge-Ampere equation
since forF(uij)
:=the matrix is definite if and
only
if the same is true for(uij).
An attempt to
give
a better definition is made in[6]65
where theequation
is called
elliptic
on agiven
solution u if at anypoint
x E Q the matrix withentries
(D2u (x))
isnonnegative (or nonpositive).
After thatequation (2.8)
is called
elliptic
in agiven
class C of functions(say,
C isC2 (S2)
or the set ofall smooth convex
functions)
if it iselliptic
on any(if
there isany)
solutionu E C. It is worth
noting
thatonly
in rare cases we can take C = in this definition. Forinstance,
as we have seenabove,
this is notpossible
for theMonge-Ampere equation. However,
theMonge-Ampere equation
iselliptic
onconvex functions. But how to find an
appropriate
C for thefollowing equations:
If we are
only
interested in definiteness of(7~..),
then as easy tocheck,
equations (2.9), (2.10)
are bothelliptic
in the same class of functions C definedas the set of all functions for which
1/,JÏ8.
It turns out that ingeneral
the Dirichlet
problem
for(2.9)
is solvable in this class and for(2.10)
is not, andmoreover the behavior of solutions of
(2.10)
is such that thisequation
shouldnot be called
elliptic
at all.Other flaws of the definition are also related to the fact that the
objective
is not
only
togive
a definition of nonlinearelliptic equation,
but to find sucha definition which could do the
job.
Forinstance, usually
we are interested inproving uniqueness,
andusually
we prove it via the maximumprinciple.
Inother
words,
if we aregiven
two solutions ui, u2 ofequation (2.8),
thenby proceeding
as usual(cf.,
forinstance, [22]62
Ch.4,
Section6.2)
for v = u 1- u2 we writewhere
and we expect the matrix a = to be
positive
ornegative.
If we assumethe above definition from
[6]65,
then we know that the matrices are say,positive
on u 1 and on u 2, butgenerally speaking,
tul 1 -f-(I - t ) u 2
is not a solution and we do not knowanything
about definiteness of a.Actually,
it mayeven
happen
that for one function F the matrix a isalways positive,
and foranother function
F, defining
anequation equivalent
to the initialequation (2.8),
the
corresponding
matrix a is neitherpositive
nornegative.
Thepoint
is thatwe can
arbitrarily modify
the function F outside the setr,
theonly
set wheresome
properties
of F aregiven
so far.By
the way, thispossibility
ofmodifying
nonlinear
equations
is the main reason for the radical difference between linear and nonlinearequations,
since for the linear case the set r is ahyperplane
inthe linear space
where k = and there are not so many ways to
represent
ahyperplane
asnull set of a linear function.
One way to overcome the last
difficulty
is toaccept
the notion ofelliptic convexity
of F from[6]65,
that is to consideronly
F such that for any twosolutions
(from
the classC)
the matrix a ispositive.
In this system ofnotions, given
anequation,
to decide if it is a"legal" elliptic equation,
we first shouldguess in what class of functions we will look for solutions and then to
modify (if
it ispossible
atall)
the functionF,
withoutchanging
theequation,
in orderto
replace
it with anelliptically
convex F. For theMonge-Ampere equations
appropriate
modifications areUnfortunately,
even after this other difficulties still remain. Forinstance,
assume that at the very
beginning
we know theappropriate
class of functionsC,
and our F iselliptically
convex in this class. Assume that we even obtaineda
priori
estimates for solutions of theequation.
Thequestion
arises how toprove existence theorems.
Usually
we introduce a parameter t E[0, 1]
and we try to find functionsFt
continuous in t such thatFI =
F andFo
defines anequation
for which every-thing
is known. After this we aretrying
to prove the same apriori
estimates forsolutions, belonging
to the same classC,
of theequations corresponding
toFt
for all t E[0, 1],
and then weapply
sometopological
methods to get thesolvability
of the
equation
= 0 for t = 1 from itssolvability
for t = 0. But onthis way, in all
interesting
cases, we cannot afford to takeFo
linear sinceusually
solutions of linear
equations
have no reasons tobelong
to C. Forinstance,
forthe
Monge-Ampere equation
det = 1 in astrictly
convex domain S2 withboundary
data on8Q,
one of theright
classes of solutions is the class of allconvex functions. At the same time there is no linear
equations
for which all solutions with differentboundary
data are convex.In a way, this cuts us off the linear
theory
and raises the obscureproblem
offinding
"model" nonlinear functionFo
for anyparticular
F. Forprofessionals
in the field this
problem
is not toohard,
and many authorsprefer
to use modelequations
whiletreating
concreteequations (see,
forinstance,
Bakel’man[6]65,
Caffarelli, Nirenberg
andSpruck [14]84,
Ivochkina[35]89 ).
But for a"ready-to-
use"
theory
this "cut off" ishighly
undesirable sinceapplications
may advanceequations
different from those which havealready
beeninvestigated. However,
in the above system of notions we cannot avoid this
difficulty
unless we caneither understand how to
modify
the method ofcontinuity
in the situation whenthe set
Ct
of solutions isevolving
with t, or we can "hide" the set Cby finding
afunction
F
such that any solutionof (2.8) of class
C is a solution to theequation
and vice versa, any solution of
(2.12)
is a solution of(2.8)
andbelongs
to C.Our definition is based on the latter
possibility.
Following [50]95
we shall present a differentapproach
to the notion ofnonlinear
elliptic equation.
We shallgive
a method to decide if agiven
nonlinearequation
is anelliptic
oneby looking only
at theequation
withoutusing
any informationregarding
theproblem
in which thisequation appeared.
After thiswe
give
a notion of admissible solutions of theequation
and then we discussthe
possibility
ofrewriting
theequation
with thehelp
ofelliptically
convexfunctions F.
The most
important
concept in ourapproach
is the notion of admissible solutions which shows theright
class of functions in which to look for solutions.This notion is based on the notion of
elliptic
branches of thegiven equation,
which turns out to be
meaningful
even forviscosity
solutions of thefirst
ordernonlinear
equations.
It is worth
noting
that in all cases known from the literature our class of admissible solutions coincides with the known ones.Also,
our notion has manycommon features with similar notions or
hypotheses
fromCaffarelli, Nirenberg
and
Spruck [ 16] g5 , Trudinger [84]90.
Our
point
of view is based on the observation that every individual equa- tion(2.8)
means and meansonly
that for any x E S2This
point
of view allows us to concentrate onproperties
of the set r ratherthan occasional
properties
of numerous functions which define the same set r.Only properties
of the set r define the type of theequation.
Of course, we assume that F is at least a continuous
function,
whatimplies
that r is a closed set in the linear space
Sd.
We alsokeep
theassumption
thatr #
0.Finally,
remember that I is the unit d x d matrix.DEFINITION 2.1. We say that a nonempty open
(in
sete =1= Sd
is a(positive) elliptic
set if(a) @ = @ ) 8 (@),
(b)
for any E30, ~
ER d
it holds that(Uij +~~)
E 8.DEFINITION 2.2. We say that
equation (2.8) (or,
moregenerally,
equa- tion(2.13)
with any nonempty closedr)
is anelliptic equation
if there isan
elliptic
set 0 such that a0 C r. In this case we call theequation
an
elliptic
branchof equation (2.8) (or (2.13)) defined bye.
Nonlinear
equations
may have manyelliptic
branches. For instance(2.9)
has two and
(2.11)
has fourelliptic
branches.DEFINITION 2.3. We say that an
elliptic
set 0 isquasi nondegenerate
if forany E
a O, ~
ER d B 101
we have -f-~’~i)
E 0.Given a number 3 >
0,
we call anelliptic
set 66-nondegenerate (or
uni-formly elliptic)
if for any W Ea O ,
ERdwe
haveIf
equation (2.14)
is anelliptic
branch of(2.8) (or (2.13))
and 6 isquasi nondegenerate (8-nondegenerate, uniformly elliptic),
we call this branch andequation (2.8) (or (2.13))
itselfquasi nondegenerate (respectively, 3-nondegener-
ate,
uniformly elliptic).
Notice that each of two
elliptic
branches of(2.9)
isuniformly elliptic
whereasall branches of
(2.11)
aredegenerate.
DEFINITION 2.4. Given an
elliptic equation (2.8) (or (2.13)),
we say that a function u is an admissible solution in Q if u is a solution in Q of anyelliptic
branch of the
equation (the
branch should be the same in the whole ofQ).
Note, for
instance,
thatu(x, y) = x2 - y2
is not an admissible solution of theelliptic equation uxx u yy
= 16.The
following
theorem shows thatequations
written in somewhat unusual form(2.14)
areactually
theequations
which one treats in thegeneral theory
offully
nonlinearelliptic equations. Exactly
this theoremjustifies
our definition.THEOREM 2.1. Let O be an
elliptic
set andequation (2.14)
beelliptic ( for
instance, be an
elliptic
branchof (2. 8) ). Define
Then
and in
particular, equation (2.12)
isequivalent
toequation (2.14). Furthermore, for any ~
E ESd
Moreover, the
function F
iselliptically
convex in the sense thatfor
any(uij),
E
Sd
thedifference F (u ij) -
can bevij)
witha
nonnegative symmetric
matrix a.Finally, if equation (2.14)
is3-nondegenerate,
then
- - .. -
An immediate consequence of this theorem and of results from
Crandall,
Ishii and Lions[23]92
is thefollowing
THEOREM 2.2. Let S2 be a bounded smooth
domain, and 0
be a continuousfunction
on a S2. Assume thatequation (2.8)
has auniformly elliptic
branch. Then thisequation
with theboundary
condition u= 0
on a S2 has an admissibleviscosity
solution u E
C (S2).
Moreover, everyuniformly elliptic
branchof (2. 8)
has its ownunique
admissibleviscosity
solution u EC(Q).
°
One of the hardest and
exciting
openproblems
in thegeneral theory
offully
nonlinear
elliptic equations
concerns smoothness of solutions when neither 6nor its
complement
is convex. If d >3, nothing
is known about boundednessor
continuity
of second order derivatives of solutions. Forexample, nothing
is known about classical
solvability
of the Dirichletproblem
for thefollowing equation
where 1 k d. Theorem 2.2
only
says that the Dirichletproblem
isuniquely
solvable in the class of
viscosity
solutions.Note that in Theorem 2.1 the
function F
isobviously
concave if 6 is convex,and it is convex if the
complement
of 0 is convex.Graphs
of convex or concavefunctions can be
represented
asenvelopes
of theirtangent planes.
Thereforeequation (2.12)
can be rewritten in the form of Bellman’sequation (1.3).
Ac-tually,
as easy to see even ingeneral
caseequation (2.12)
isequivalent
to aBellman
equation,
which contains sup and inf at the same time. If we combine this with results from[44]85,
then we obtainTHEOREM 2.3. Let Q be a bounded domain
of
classC2+,
where a E(0, 1),
andlet 0
EC2+"
Assume thatequation (2.8)
has auniformly elliptic
branchdefined by
a domain 6 such that either 6 or itscomplement
is convex. Then thisequation
with theboundary
conditionu - ~
on a S2 has an admissible solutionu E
c2+fJ (Q),
wheref3
E(0, 1 ).
Moreover, theelliptic
branch(2.14)
with thegiven boundary
condition has its ownunique
admissible solution u EC2+0
This theorem
applies
toequation (2.9)
which has twouniformly elliptic
branches.
General
theory
from[46]89
or[51 ]95
alsoimplies
thefollowing
theoremwhich can be restated in an obvious way for the case in which the
complement
of E) is convex.
THEOREM 2.4. Let E) be an open convex set and let
equation (2.14)
beelliptic.
Let C be an open cone in
Sd
with vertex at theorigin,
and let to be a number. Assume thatto I
+ E) CC,
andthat for
any w E C we have tw E 0for
all tlarge enough.
Let tr w > 0
for
anyw E C,
and let Q be astrictly
convex domainof
classC4.
Then
for any 0
E there is aunique function
ú EC (0)
n(Q)
such thatu E ae
(a.e.)
in Q.If,
in addition,equation (2.14)
isquasi nondegenerate,
then u EC2+" (0) for
an a E(0, 1).
This is not the most
general
result from[46]89
or[51]95.
There we do notassume that Q is convex or that the weak
nondegeneracy
condition: tr w > 0 for anyw E C,
is satisfied. The assertion then becomes morecomplicated.
The
general
theoremapplies
to each of fourelliptic
branches of(2.11 ).
Forthe branch = 0 it says that for any
C4
functiongiven
on theboundary
of aC4
domain Q CJR2
there exists aunique C1,I(Q)
solution with thisboundary
valueprovided
that theboundary
of Q isstrictly
convex(outward)
at any
point
where thetangent
line isparallel
to one of the coordinate axis.We say more about these results in Section 4.
To conclude this section we mention a
general
method how toperturb
anelliptic equation (2.8)
in order to get auniformly nondegenerate
one. It sufficesto take s > 0 and consider
This method has been
constantly
in use in worksby
the author.Usually,
inthe literature other methods are
applied.
In connection with this it is worthnoticing
that thefollowing
"natural"perturbation AM + det D2U
= 1 of theMonge-Ampere equation detd 2U
= 1 is notelliptic
at allunless s s
0 !3. - General convex
fully
nonlinearuniformly elliptic equations
Since works of Bernstein it was known that to prove the
solvability
of dif-ferential
equations
it suffices to obtain apriori
estimates inappropriate
classesof functions for solutions under the
hypothesis
that the solutions exist. Oneof ways to obtain estimates in is to differentiate the
equation
once thusobtaining
aquasilinear equation
or a system ofequations
with respect to first derivatives andhope
that there isCl+"-estimate
for solutions of this new equa- tion. Thishope
wasdestroyed by
Safonov[77]87 .
Theexample
in[77]g~
buriedthe
hopes
for an"easy" theory
offully
nonlinearequations,
and in a sense savedthe work which has been done before for F’s either convex or concave with respect to the matrix
D 2u
andsufficiently
smooth in other variables. Below weonly speak
about such F’s.In works
by
Evans[25 ] g2
and the author[40]82
we obtained interior apriori
estimates. Then in[41 ]83
the authorpublished
apriori
estimates up to theboundary
for the Dirichletproblem.
After this many authors con-tributed to the
general theory
ofuniformly elliptic equations.
We willonly
mention works which
played
the mostimportant
role in thedevelopment
ofthe
general theory.
We start with worksby
Evans[26]83 , Trudinger [8 1]83, Krylov [43]g4, Caffarelli, Kohn, Nirenberg
andSpruck [15]85, Caffarelli, Nirenberg, Spruck [ 16]gs .
The
major
results obtained before 1984 are summed up in the booksby Gilbarg
andTrudinger [29]g3
andKrylov [44]85 .
After thebreakthrough
madein the papers
by
Evans and the author usualtechnique
allowed todevelop
atheory
which contains thegeneral theory
ofquasilinear elliptic
andparabolic equations.
Inparticular,
the famousLadyzhenskaya-Ural’tseva
theorem has beengeneralized
forfully
nonlinearequations.
It is to be mentioned that in all theseworks the data were assumed to be smoother than in linear
theory,
so that ifequation ( 1.1 )
isjust
Au +f
= 0 and we want toget
itssolvability
from thegeneral theory
offully
nonlinearequations,
then we should assume thatf
issmooth
enough.
A
major
step forward in thegeneral theory
has been doneby
Safonov in 1984 whoby using
anentirely
newtechnique proved
thesolvability
inC2+,
for(1.1)
under
only
natural smoothnessassumptions
on F(see [76]84
and[78]88 ).
Thisis an
extremely
strong result which issharp
even for linearequations.
Whatis even more
surprising,
Safonov’sproof
of estimates for(1.1)
goes thesame way when the
equation
islinear,
and in this very case it is much easier and shorter than the knownproofs
of these estimates for linearequations.
Wepresent
thisproof
in Section 7.The above mentioned works deal with the Holder space
theory
offully
non-linear
elliptic equations.
The firstbreakthrough
in the Sobolev spacetheory
hasbeen done
by
Caffarelli[9]g9 (also
see the bookby
Caffarelli and Cabre[13]95)
for the
elliptic
case andWang [98]92
for theparabolic equations.
The worksby Safonov,
Caffarelli andWang
are remarkable in one morerespect-they
donot suppose that F is convex or concave in
D2u.
But in thegeneral
casethey only
show that to prove apriori
estimates it suffices to prove the interiorC2-estimates
for "harmonic" functions.So far we were
talking
about the Dirichletproblem.
Nonlinearoblique
derivative
problems
wereinvestigated
as well. Anexample
of such conditionsis the
following capillarity boundary
conditionThe most relevant references here are Lions and
Trudinger [63]85 ,
Libermanand
Trudinger [60]86 ,
Anulova and Safonov[5]86 . Again
as in the case of theDirichlet
problem,
the results forfully
nonlinearequations
contain those forquasilinear equations.
4. - General
degenerate fully
nonlinearelliptic equation
Fully
nonlineardegenerate elliptic equations
arise inapplications
much moreoften than
uniformly elliptic
ones,though
theirinvestigation
in classesC2+a heavily
relies on results from thetheory
of nonlinearuniformly elliptic equations.
There is a substantial difference in difficulties which arise when we are
dealing
withdegenerate equations
in the whole space or in bounded domains.The
theory
in the whole space has beendeveloped mostly by probabilistic
meansand is understood to a very
good
extent.Many
mathematicians contributed tothe
probabilistic
version of thetheory,
between them are Lions and the author.A PDE counterpart of this
theory
can be found in[44]85 . Degenerate equations
are
important
notonly
from thepoint
of view ofapplications.
Thefollowing degenerate Monge-Ampere equation
in a sense, even
plays
the main role in thetheory
ofuniformly elliptic equations.
By
the way, as we haveexplained
above theequation
is
uniformly elliptic
for any E > 0.The
theory
of nonlineardegenerate equations
cannot be easier than thetheory
of linear ones. It is worth
mentioning
that even in the lineartheory
there arestill very many unsolved
problems. Probably
the best referencesconcerning
thelinear
theory
areKohn, Nirenberg [38]6~
andOleinik,
Radkevich[68]71.
After the book
[44]85
the firstgeneral
results onfully
nonlineardegener-
ate
equations
in domains were obtained in 1985by Caffarelli, Nirenberg
andSpruck [16]85,
wherethey
consideredequations
likewhere *
is agiven
function of x,f
is asymmetric
function of h ER d
andh(u)
= is a vector ofeigenvalues
ofEquation (1.7)
is aparticular
case of(4.15).
In[181"
the authorsapply
theirtheory
to curvaturesof the
graph
of u instead ofh(u)
and prove thesolvability
ofequations
forstarshaped
surfaces. It is worthnoticing
thatequations
with curvatures aremuch more
complicated
thancontaining only D2u.
One can say that the latterare
linearly fully
nonlinearequations
whereas the former arequasilinearly fully
nonlinear
equations.
In both cases one deals withequation
like(1.3)
but in thecase of
linearly fully
nonlineartheory
one assumes that a isindependent
ofDu, u and b is linear at least with respect to Du.
A
general
theorem has been announcedby
the author in 1986(see [45]g6)
and
proved
in PDE terms in[48]93
and[51 ]95 .
One of its versions ispresented
above in Section 2. This theorem allows one to consider a very
large
class of(linearly) fully
nonlineardegenerate elliptic equations.
Between them are theequations
The first one is the heat
equation
which is aparticular
case ofdegenerate fully
nonlinear
elliptic equations.
Thegeneral
theoremimplies
that if S2 ={(t, x)
E+ r2 }
and r(3 - ~)d,
then for anyC4-function
g on theboundary
there exists aunique
solution u E such that u = g on a 0.This result turned out to be unknown in the
theory
of lineardegenerate equations.
Also,
in what concerns the restrictions on r and g, it issharp
asWeinberger’s example
from[38]6~
shows. For the secondequation
the theorem says that in anyC4 strictly
convex domain Q and anyC4-function
g theequation
has aunique
convex solution of class 1 such that u = g on Theexamples
fromCaffarelli, Nirenberg
andSpruck [17]g6
show that all the above conditions arenecessary. In
particular,
anexample by
Urbas shows thatgenerally
the solutionis not better than 1 even for
analytic boundary
data in a ball. The sameis true for the third
equation only
we have tospeak
aboutplurisubharmonic
uand
strictly pseudoconvex
domain Q. Theexample by
Urbas admits an easycomplexification.
One more
example
ofapplications
of thegeneral
theorem from[48] 93
and
[51 ]95
is thefollowing equation
where
Pk (A)
are the k-thelementary symmetric polynomials
ofeigenvalues
ofthe matrix A.
As in the case of linear
theory,
thetheory
of nonlineardegenerate equations
is rather far from
being
welldeveloped.
One of veryimportant questions
isinterior
regularity
of solutions. Thepoint
is that in manyexamples
ofequations
of